Google发布Gemini 4 Hydrogen:主打Agent工作流
Google Releases Gemini 4 Hydrogen: 2.5x Faster Than Gemini 3.8 Flash
Gemini 4 Hydrogen是Google针对Agent场景推出的全新旗舰级专用模型,不仅刷新了多项基准测试成绩,更提供了明确的模型分工策略(Hydrogen执行+Carbon规划),对构建多Agent系统的开发者极具参考价值。
Google has released Gemini 4 Hydrogen, a fast model for coding and agentic work. Gemini 4 Hydrogen is a proprietary model from Google DeepMind built for everyday agent loops. It sits below the frontier models Gemini 4 Carbon and Gemini 4 Argon. Google calls it its “daily driver for agentic workflows.”
Google 发布了 Gemini 4 Hydrogen,这是一款专为编码和智能体工作设计的快速模型。Gemini 4 Hydrogen 是 Google DeepMind 的专有模型,旨在用于日常智能体循环。它位于前沿模型 Gemini 4 Carbon 和 Gemini 4 Argon 之下。Google 称其为“智能体工作流的日常主力工具。”
TL;DR
TL;DR(太长不看版)
- Size: 1,048,576-token unified window for input, thinking and output.
- Runs on: Google-hosted API only. No weights, no quantized variants, no self-hosting. Serving hardware not disclosed.
- Performance: Scores 51 on the Artificial Analysis Intelligence Index at high, above Argon’s 50.
- Best: 57% on Terminal-Bench 4.1, versus 18.2% for Gemini 3.8 Flash.
- Worst: Ranks only #5 on Agent Arena. Costs 2x Gemini 3.8 Flash’s current introductory price.
- Bottom line: Best: frontier-class agent scores at $1.50/$7.50 per 1M tokens. Worst: closed, and synchronous output caps at 65,536 tokens.
- 规模:输入、思考与输出采用统一的 1,048,576 token 窗口。
- 运行环境:仅限 Google 托管 API。无权重、无量化变体、不支持自托管。未披露服务硬件。
- 性能:在 Artificial Analysis Intelligence Index 中,高设置下得分为 51,高于 Argon 的 50。
- 最佳表现:在 Terminal-Bench 4.1 上得分 57%,而 Gemini 3.8 Flash 仅为 18.2%。
- 最差表现:在 Agent Arena 中仅排名第 5。成本为 Gemini 3.8 Flash 当前入门价格的 2 倍。
- 总结:最佳——以每百万 token 1.50/7.50 美元的价格提供前沿级智能体评分。最差——封闭模型,且同步输出上限为 65,536 token。
What is Gemini 4 Hydrogen?
什么是 Gemini 4 Hydrogen?
Hydrogen is a closed-weights model served through the Gemini API and Google products. The model card says it is adapted from Gemini 4 Carbon. Google does not disclose parameter count, architecture or training hardware. The knowledge cutoff is March 2026.
Hydrogen 是一个闭源权重模型,通过 Gemini API 和 Google 产品提供服务。模型卡片显示其基于 Gemini 4 Carbon 调整而来。Google 未披露参数量、架构或训练硬件。知识截止日期为 2026 年 3 月。
Input covers text, images, audio, video, PDFs, tables and charts, per the What’s New guide. Output is text. The 1M window is shared by input, thinking and output tokens.
根据《新功能》指南,输入涵盖文本、图像、音频、视频、PDF、表格和图表。输出为文本。1M 窗口由输入、思考和输出 token 共享。
How do the 4 thinking levels work?
四种思考级别如何运作?
Developers set thinking_level inside generation_config. Values are low, medium, high and a new xhigh. The default is high, not medium as on 3.8 Flash. A minimal setting now returns an error. The model card warns that higher levels use more tokens.
开发者在 generation_config 中设置 thinking_level。可选值为 low、medium、high 以及新增的 xhigh。默认设置为 high,而非 3.8 Flash 中的 medium。现在若设置为最小值将返回错误。模型卡片警告更高级别会消耗更多 token。
How does Hydrogen work with Carbon and Argon?
Hydrogen 如何与 Carbon 和 Argon 配合工作?
The key design idea is role-splitting across models. In Google Antigravity, an Orchestrator option switches models inside 1 task. Carbon explores, resolves ambiguity and plans. Hydrogen edits files, updates packages and runs tests. When Hydrogen hits friction, it hands control back to Carbon.
核心设计理念是跨模型的角色分工。在 Google Antigravity 中,Orchestrator 选项可在单个任务内切换模型。Carbon 负责探索、消除歧义和规划。Hydrogen 负责编辑文件、更新包并运行测试。当 Hydrogen 遇到阻力时,会将控制权交还给 Carbon。
On Terminal-Bench 4.1, this pairing reached near-Carbon task success rates. It used 24% less quota and finished 19% faster than Carbon alone. The Antigravity SDK exposes the same split for custom multi-agent systems.
在 Terminal-Bench 4.1 上,这种配对达到了接近 Carbon 单独运行时的任务成功率。相比单独使用 Carbon,配额使用减少了 24%,完成速度提高了 19%。Antigravity SDK 为自定义多智能体系统提供了相同的拆分机制。
How does Gemini 4 Hydrogen perform on benchmarks?
Gemini 4 Hydrogen 在基准测试中的表现如何?
All figures are self-reported by Google unless noted.
除非另有说明,所有数据均由 Google 自行报告。
- Artificial Analysis Intelligence Index: 51 at high, 54 at xhigh. Argon scores 50 on Index v5.0.
- Terminal-Bench 4.1: 57%, versus 18.2% for 3.8 Flash.
- HLE-Diamond: 58.8%, leading its class.
- Vals Tax Agent Bench: 52.3%, ranked #1 overall.
- Vals Legal Research Bench: 55.8%, tied for #1.
- Multimodal Knowledge Agent: 53.4%.
- Agents’ Last Exam v1.1: 71.6%.
- Agent Arena: ranked #5.
- Artificial Analysis Intelligence Index:高设置下得分为 51,xhigh 设置下为 54。Argon 在 Index v5.0 中得分为 50。
- Terminal-Bench 4.1:57%,而 3.8 Flash 为 18.2%。
- HLE-Diamond:58.8%,在其类别中领先。
- Vals Tax Agent Bench:52.3%,总体排名第 1。
- Vals 法律研究基准:55.8%,并列第一。
- 多模态知识智能体:53.4%。
- 智能体终极考试 v1.1:71.6%。
- 智能体竞技场:排名第 5。
Speed is the headline claim. Hydrogen finishes tasks 2.5x faster than 3.8 Flash at the default high level. In partner tests, it ran 2x to 2.8x faster with 12% to 23% fewer tokens.
速度是主要卖点。在默认高级别下,Hydrogen 完成任务的速度比 3.8 Flash 快 2.5 倍。在合作伙伴测试中,其运行速度快了 2 到 2.8 倍,且使用的 token 数减少了 12% 至 23%。
Google also reports more than 5x fewer fabricated claims than 3.8 Flash. It scored 96.7% on an internal agent honesty evaluation. It was 2.3x more resilient to indirect prompt injection on Gray Swan’s benchmark.
Google 还报告称,其捏造声明的数量比 3.8 Flash 少 5 倍以上。它在内部智能体诚实度评估中得分 96.7%。在 Gray Swan 的基准测试中,它对间接提示注入的抵抗力强了 2.3 倍。
How much does Gemini 4 Hydrogen cost?
Gemini 4 Hydrogen 的价格是多少?
The API price is $1.50 per 1M input tokens and $7.50 per 1M output tokens. Pricing is flat across the full window. Output pricing includes thinking tokens. Implicit caching cuts input to $0.075 per 1M for prompts of 4,096 tokens or more. Explicit caching is not supported.
API 价格为每 100 万输入 token 1.50 美元,每 100 万输出 token 7.50 美元。定价在整个上下文窗口内保持一致。输出价格包含思考 token。隐式缓存将 4,096 个 token 或以上的提示的输入价格降至每 100 万 token 0.075 美元。不支持显式缓存。
The price context matters. Gemini 3.8 Flash costs $0.75/$3.75 until December 31, 2026. From January 1, 2027, it rises to $1.50/$7.50, matching Hydrogen. Argon costs $2/$10 until November 30, 2026, then $4/$20.
价格背景很重要。Gemini 3.8 Flash 的价格为 $0.75/$3.75,有效期至 2026 年 12 月 31 日。从 2027 年 1 月 1 日起,价格将上涨至 $1.50/$7.50,与 Hydrogen 持平。Argon 的价格为 $2/$10,有效期至 2026 年 11 月 30 日,之后为 $4/$20。
Where can developers use Gemini 4 Hydrogen?
开发者可以在哪里使用 Gemini 4 Hydrogen?
The model ID is gemini-4-hydrogen. It is live in the Gemini API, Google AI Studio on all tiers, and Android Studio. It also ships in Gemini for business, the Gemini app, Google Search and Sheets. Antigravity access requires Google AI Pro or Ultra. Vertex AI is not mentioned in the launch posts.
模型 ID 为 gemini-4-hydrogen。它已在 Gemini API、所有层级的 Google AI Studio 以及 Android Studio 中上线。它也随附于面向企业的 Gemini、Gemini 应用、Google 搜索和 Sheets。Antigravity 访问需要 Google AI Pro 或 Ultra 套餐。发布帖子中未提及 Vertex AI。
Migration from 3.8 Flash needs a few changes. Remove temperature, top_p and top_k, since the server ignores them. Remove cached_content fields. Raise hardcoded max_output_tokens caps to 65,536.
从 3.8 Flash 迁移需要进行一些更改。移除 temperature、top_p 和 top_k,因为服务器会忽略它们。移除 cached_content 字段。将硬编码的 max_output_tokens 上限提高至 65,536。
Gemini 4 Hydrogen vs competitors
Gemini 4 Hydrogen 与竞争对手对比
| Feature | Gemini 4 Hydrogen | Gemini 3.8 Flash | Gemini 4 Argon | Claude Haiku 5.5 |
|---|---|---|---|---|
| Release | Nov 12, 2026 | Sep 2, 2026 | Oct 13, 2026 | Oct 7, 2026 |
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Context | 1,048,576 unified | Not stated in launch post | 1M unified | Not stated on launch page |
| Weights | Closed | Closed | Closed | Closed |
| Input $/1M | $1.50 | $0.75 (intro), $1.50 from Jan 1, 2027 | $2 (intro), $4 from Dec 1, 2026 | $0.10 (≤100K), $0.50 (>100K) |
| Output $/1M | $7.50 | $3.75 (intro), $7.50 from Jan 1, 2027 | $10 (intro), $20 from Dec 1, 2026 | $0.50 (≤100K), $2.50 (>100K) |
| AA Intelligence Index | 51 (high), 54 (xhigh), per Google | Not disclosed | 50 (v5.0) | 36 (v5.0.1) |
| Terminal-Bench | 57% (4.1) | 18.2% (4.1, per Google) | 57% (4.0, AA) | 39.2% (4.0, max effort) |
| Reasoning control | 4 thinking levels | Effort levels | Not stated | Effort settings |
| 功能 | Gemini 4 Hydrogen | Gemini 3.8 Flash | Gemini 4 Argon | Claude Haiku 5.5 |
|---|---|---|---|---|
| 发布日期 | 2026 年 11 月 12 日 | 2026 年 9 月 2 日 | 2026 年 10 月 13 日 | 2026 年 10 月 7 日 |
| 参数量 | 未披露 | 未披露 | 未披露 | 未披露 |
| 上下文窗口 | 1,048,576 统一上下文 | 发布帖中未说明 | 1M 统一上下文 | 发布页面未说明 |
| 权重 | 闭源 | 闭源 | 闭源 | 闭源 |
| 输入价格$/1M | $1.50 | $0.75(促销),2027 年 1 月 1 日起 $1.50 | $2(促销),2026 年 12 月 1 日起 $4 | $0.10(≤100K),$0.50(>100K) |
| 输出价格$/1M | $7.50 | $3.75(促销),2027 年 1 月 1 日起 $7.50 | $10(促销),2026 年 12 月 1 日起 $20 | $0.50(≤100K),$2.50(>100K) |
| AA 智能指数 | 51(高),54(极高),据 Google 称 | 未披露 | 50(v5.0) | 36(v5.0.1) |
| Terminal-Bench | 57% (4.1) | 18.2% (4.1,据 Google 称) | 57% (4.0,AA) | 39.2% (4.0,最大努力) |
| 推理控制 | 4 个思考级别 | 努力级别 | 未说明 | 努力设置 |
Terminal-Bench 4.0 and 4.1 are different versions, so cross-version scores are not directly comparable.
Terminal-Bench 4.0 和 4.1 是不同的版本,因此跨版本分数不可直接比较。
Key Takeaways
关键要点
- Hydrogen scores 51 on the AA Index, above Argon’s 50, per Google.
- It hits 57% on Terminal-Bench 4.1, versus 18.2% for 3.8 Flash.
- Pricing is $1.50/$7.50 per 1M tokens, flat across 1M context.
- Default thinking is now high, with a new xhigh tier.
- Pair it with Carbon in Antigravity for 24% lower quota use.
- 据 Google 数据,Hydrogen 在 AA Index 上得分为 51,高于 Argon 的 50。
- 它在 Terminal-Bench 4.1 上得分达到 57%,而 3.8 Flash 仅为 18.2%。
- 定价为每百万令牌 1.50 美元/7.50 美元,在 1M 上下文范围内价格不变。
- 默认思考模式现已设为高,并新增 xhigh 层级。
- 将其与 Antigravity 中的 Carbon 搭配使用,可降低 24% 的配额消耗。
Check out the launch post, the model card, the developer post and the What’s New guide. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
请参阅发布文章、模型卡片、开发者帖子以及“更新内容”指南。此外,欢迎在 Twitter 上关注我们,别忘了加入拥有 15 万+成员的 ML SubReddit,并订阅我们的新闻通讯。等等!你在 Telegram 上吗?现在你也可以加入我们的 Telegram 群组。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力